Abstract:

Objective: To address the challenges of low identification accuracy and insufficient robustness in existing detection algorithms for alien forest pests, which stem from factors such as diverse target scales, complex habitats, and frequent occlusions, this paper proposes a detection method integrating multi-scale attention features, termed MAF-YOLO. The method aims to significantly enhance the detection accuracy and generalization capability for alien pests in complex natural environments by strengthening the model’s feature extraction capability for targets of varying scales, particularly small-scale ones, and optimize the bounding box regression strategy. Method: The proposed method was based on the YOLOv5s baseline architecture, and its core modifications included: 1) A coordinate attention (CA) mechanism was embedded into the neck network to enhance the capture of key target features and suppress background noise. 2) A small target detection head with 160×160 pixel was added to construct a multi-level detection structure, thereby improving detection sensitivity for minute objects. 3) The Focal-EIoU loss function was employed to replace the original CIoU loss, in order to mitigate the imbalance between positive/negative and easy/hard samples, and refine object localization accuracy. 4) A domain adaptation training strategy was introduced to improve the model's generalization across diverse scenarios by pre-training on a large-scale general-purpose dataset. Result: The proposed model was trained and evaluated on an image dataset comprising 15 categories of potential and existing invasive forest pests. The improved YOLOv5s model, MAF-YOLO, demonstrated an increase in precision and recall by 3.6% and 4.4%, respectively, compared to the original YOLOv5s. In comparison with the SSD, YOLOv7, and YOLOv8 models, the average precision of the improved model was higher by 2.2%, 1.1%, and 0.3%, respectively. Furthermore, with the integration of domain adaptation, the model’s precision reached 77.9%, representing a 2.1% improvement over the baseline model under the same strategy. Conclusion: This study achieves precise recognition of invasive alien forest pests, enhances the model's accuracy and robustness, and provides theoretical and technical support for effective monitoring of alien forest pest.

Key words: alien forest pest, YOLOv5s, coordinate attention, domain adaptation, image detection